Accurate recognition of multiple events for DAS

Fiber-optic distributed acoustic sensors (DAS) have been widely employed in urban safety monitoring, seismic activity detection, and related fields. However, in multi-scene applications, recognition accuracy often deteriorates due to event diversity and environmental variations. Additionally, training models for new scenes is time-consuming and labor-intensive, posing significant challenges for efficient and accurate event recognition. This paper presents a novel method for accurate recognition of multiple events across scenarios using DAS. The proposed approach employs an encoder-LSTM network to extract robust event features and utilizes spectral clustering to construct a hierarchical classification tree. By recursively grouping event features into leaf nodes based on similarity, a complete classification structure is formed. The method integrates multi-task learning for simultaneous training of all tree nodes and incorporates an incremental learning mechanism to enable efficient updates for new event types. Field tests validate the effectiveness of this method across three scenarios with nine event types, achieving a recognition accuracy of 96.60%, outperforming traditional multi-classifier approaches. When a new event type is added, the incremental learning model maintains an accuracy of 96.55% for ten event types while reducing training time to 1.5% of the time required by conventional methods.